Gemini Live still takes visual context in the same conversation as speech, so if the product sees, consent has to cover faces, screens and rooms – not only the soundtrack.
Google’s own product pages have not walked this back. You start a Live chat with your voice. You tap camera and show Gemini what you see. You share a screen and keep talking. You can switch between talking and typing without opening a new thread. Context, tools and history stay in one conversation. Camera sharing can resume after a pause. Screen sharing has to be restarted, which is a UX footnote, not a rights boundary. There is on-screen visual guidance that highlights the object the model has decided you meant. There is a generate-from-what-you-see path, the “reimagine this” trick, sitting on the same feed. Interruptions are a setting. Captions are a setting. The camera is not a setting you can treat as audio-only after the fact.
We already wrote that live agents can see, and that is a faces problem until you say otherwise. The news this month is that the product did not split. Speech-to-speech stacks at OpenAI and Google still sell one conversation that hears and looks. Enterprise copies are in shops, in field-service apps, in “show me the boiler” helpdesks. If you briefed audio last year and turned the camera on because Live can ground on what it sees, you do not have a feature flag. You have a second capture in the same session, and often a second consent you never took.
The product still sees in the same conversation
A year ago you could pretend the visual path was a demo. Full-duplex already forced overlap into the audio spec. Visual context forces a second object into the same session: hands, packaging, a passport, a kitchen counter, a colleague walking through, a laptop with mail open. Because it is the same conversation, the model can use what it just heard to interpret what it now sees, and what it just saw to interpret the next utterance. That is the pitch. It is also why you cannot collect the soundtrack, add a webcam later, and call the pack multimodal. The behaviour you will ship is joint.
Device is part of that joint object. Phone rear camera in a kitchen. Front camera as a selfie. Screen share of a checkout flow. A kiosk in a shop. Glasses in a home. Do not call all five “video.” The environment is different. The bystanders are different. The lighting is different. Eval collected on a kiosk will not tell you whether glasses in a kitchen work. Screen share is not a camera. It is a stream of other people’s documents, order numbers, message threads and faces in thumbnails. If the product can see a screen, the brief has to name screens as in-scope or as a reject. “We only wanted the SKU” is not a crop. It is a hope.
Locale still matters, and it now sits on two channels at once. Language on the soundtrack and language on the packaging can disagree. A shop in Dublin is not a shop in Dallas. A British kitchen with a bill on the counter is a different leak from a US kitchen with a driving licence. If the product claims those rooms, the files have to contain them, and the consent has to name them.
Generation sitting on the same feed is a third product. “Reimagine this” takes the live camera and makes a new picture. Consent for a recogniser is not consent for a Live agent. Consent for a Live agent that grounds on a boiler is not consent for a generator that trains on the hallway, or that stores a still, or that produces a marketing image from a family kitchen. TTS from the same session is a fourth product. If you do not split those yes-or-no lines in the ask, you will discover them in diligence.
A familiar week: the speech set is done. Someone in product turns on camera share because Gemini Live can highlight the right spanner. The files from last year are WAV. The people who recorded them agreed to a voice product. Nobody mentioned hallways, passports, a colleague walking through, or a generate step on the same stream. Friday looks like a multimodal launch. Monday is a face-rights problem with extra audio. The camera is not a feature flag. It is a new object. Often a new consent.
Audio consent does not stretch
What consented AI training data is is a specific sentence. People were told what would be captured, what it would be used for, and they agreed in a way you can evidence later. “I talked to an assistant” is not “you may train on the video of my hallway, the people who walk through it, the screen I shared, and any stills you generate from that feed.” Internal quality monitoring is not commercial multimodal training. A privacy-policy link in an app is notice to a user of an app. It is not a pack you can show an auditor for shipping, eval, marketing stills, or a third-party licence.
Europe is not going to save you with a crawler story. The Code of Practice and the training-content summary are about works and sources at model level. A live camera in a kitchen was never “it was online.” Article 10, if you are in a high-risk box, asks whether the sets match the setting. Homes are a setting. Shops are a setting. Children in those rooms are not a diversity slide. Counsel will not treat “it was a support call” as a chain for training a visual policy on the hallway, and they will not treat a Gemini-shaped product as a reason the old WAV consent grew a camera.
If faces are not required, keep them out of the capture. Framing, crop, and “we will blur later” are processing choices. They are not retroactive consent, and blur often destroys the signal the model wanted – the hands, the SKU, the document, the boiler plate. Synthetic overlay can hide a face for a demo. It does not create a cleared training set, and it may destroy the grasp. Screens in shot are their own leak. Phones, laptops, order numbers on a monitor, a passport, a boarding pass, a WhatsApp thread. Reject codes belong in the brief: task not in frame, non-consented people, children, screens, beauty filters, black frames, generate-from-this used when you only needed grounding. Automation can flag empty files. A person still has to catch the rest.
Bystanders are not a rare edge. First-person and phone cameras see kitchens because that is where the appliance is. They also see faces in the doorway, homework on the table, medicine on the counter, a neighbour through glass. Shop kiosks see other customers. Field-service glasses see the householder, the children, the photographs on the fridge. The enrolled speaker can sign. The people behind them did not. Paying the caller, or the technician, does not clear the room.
Split eval by person if identity leaks. Do not mix the visual eval into the speech fine-tune because it was sitting in the same Live bucket. That is eval leaking into training with a camera. Gold before production. Task not in frame is a reject, not a “we will use it for audio.” If you only needed audio, you should not have recorded video. If you needed video, last year’s speech set does not become multimodal because you bought a webcam.
What has to be in the first email
Does the product see? If no, audio only, and say so in the consent as well as the spec. If yes, what exactly – product SKU, document, hands, face, screen, room? Device: phone selfie, rear camera, kiosk, glasses, screen share. Whether generation from the feed is in scope. Locale. Training, eval, shipping, stills, TTS – separately. Whether you already have a catalogue match.
Write the sentence. Bad: “Gemini-Live-like, voice plus a bit of camera.” Better: “Identify the broken heating control on a UK boiler, phone rear camera, indoor daylight, enrolled adult only, faces of others are a reject, screens are a reject, no generate step, start to finish including the moment they point at the wrong valve.” If you cannot write the sentence, you are shopping for a Live demo, not collecting a skill.
Write reject codes while everyone is still calm. Task not in frame. Non-consented people. Children. Screens. Beauty filters. Digital zoom that destroys the hands. Audio-only takes sitting in a visual pack. Visual takes sitting in an audio pack. Scripted pointing when you asked for ordinary behaviour. A suspiciously empty room when you asked for a shop. Gold items before anyone records, or the first month is the vendor learning your hallway on your budget.
Rights for each use as a yes or no line. Grounding a Live agent on a boiler is one line. Training a visual policy is another. Keeping a frozen eval is another. Stills on a launch site are another. TTS from the session is another. Generating new images from the live feed is another. A creator or a customer who tapped accept on “talk to the assistant” has not automatically agreed to the rest. Expanding a research notice after the weight exists is how teams discover they have a diligence finding with extra files.
If a catalogue set already matches the rig and the rights, look at it with the actual stack you will ship. Custom when the hallway, the kiosk, the faces decision, or the screen-share rule is yours. Starting with “we will blur later” because Gemini Live made the camera feel free is how you buy a biometric pack for a product that might still be a headset.
The models can see. They can still see in the same conversation as the voice. Your files still only hear, unless you specified the rest and asked for it. That is the gap to collect, or to refuse. Seeing is a second product. Treat it like one. Do not bolt it onto a WAV pack and hope the consent stretches. It will not. Spec the fork before the first take. Audio only, and say so. Visual, and say what is in frame. Mixing them “in case” is how you buy a faces problem you did not need. Robotics already learned this with chest cameras. Speech teams are still learning it because Live models did not split the thread.
People4.ai will not bolt a camera onto a speech brief because the API still accepts one. hello@people4.ai · Contact
FAQ
Can we reuse last year’s speech set and add Gemini-style camera share later?
Only if last year’s people agreed to this purpose and this capture. A new sensor is a new object. Often a new consent. Do not bolt video onto a WAV pack and call it multimodal training data.
If we only use the camera for grounding, do we still need face and screen rights?
Yes, if faces and screens can appear. Grounding is still capture. If they are not required, keep them out at capture and reject takes that include them. Blur later is not retroactive consent.
Does screen share need a different consent from the rear camera?
The rights questions are the same. The object is not. Screen share is other people’s documents, faces in thumbnails and order numbers. Spec it as in or out. Do not call both “visual context.”
Related reading
Live agents can see now. That is a faces problem until you say otherwise. · What is consented AI training data?